After Claude wiped my node_modules, I built this
I've been letting Claude and GPT run commands in my terminal for the last year. Most of the time it's great. The other times I lose half a day to AI mistakes: rm -rf the wrong folder, force-pushing to main, "fixing" a config by quietly overwriting it.
So I built Verlox.
Every command the AI proposes gets turned into a plan first: numbered steps, the real commands, the files each step will touch, and a risk score (low / medium / high) per step. Nothing executes until I approve. Risky things (deletes, force-pushes, production access) always stop and ask separately. Anything the AI deletes or overwrites is copied to a Recovery Vault first, restorable with one click.
A few things I'm proud of:
Eleven hosted models in one menu. Claude, GPT, Gemini, Grok, Llama, DeepSeek, Qwen. Free tier with 15 credits a day across 8 of them.
A built-in offline model. Llama 3.2 3B downloads once and then runs entirely on your machine. Free, private, no network.
Simulate mode. Dry-run any plan and see the predicted before/after diffs before running it for real.
Per-capability permissions. Set "always allow / ask every time / never" once for things like file reads, package installs, network calls, production access.
How it's different from Warp, Terax, Wave: those optimize for speed first. Verlox optimizes for keeping your work intact. Same speed as letting AI rip, with a real safety net.
Honest disclaimers:
Windows only today. macOS and Linux next.
Not code-signed yet. SmartScreen will warn on first install.
Solo project. Source-available on GitHub.
Genuinely curious: what's the AI-terminal feature you wish existed but doesn't? What would make you trust an agent more in your shell?
Replies